← Back to the dossier

7 October 2026 · Anthropic

Anthropic's smallest Claude gets cheaper and starts driving a computer

Claude Haiku 5.5 is Anthropic's cheapest and fastest small model, and on the company's numbers it now costs about 75 percent less to run than Haiku 4.5. It is the first Haiku that can be told to think harder or cheaper. On a test of using a computer, it lands far above the previous Haiku and above OpenAI's GPT-6 Luna.

By Mara Masaeva · Updated 8 October 2026

CapabilityConfirmedMore than one independent source, or a primary document.

What happened

Anthropic released Haiku 5.5 on 7 October, on its own platform and on Amazon, Google and Microsoft. The company aims it at work that runs in bulk: summaries, sorting, database questions, and as a helper next to Sonnet 5.5 or Opus 5.5 when those models write code. It also names live customer support and using a browser.

For a prompt under 100,000 tokens, which Anthropic says was about 90 percent of requests to the previous Haiku, the price is 0.10 dollars per million tokens in and 0.50 out. Above that length it is 0.50 and 2.50. Haiku 4.5 was 1 dollar in and 5 out. Anthropic's average of "about 75 percent less" mixes those two prices with a small rise in tokens per task. Under 100,000 tokens the list price is 90 percent lower. Above it, 50 percent lower.

The same day, Anthropic halved what it charges to reread cached context on Sonnet 5.5, from 0.20 dollars to 0.10 per million tokens. It says that cuts the cost of most agent work on Sonnet by about 20 percent. Max and Team subscribers also get monthly API credit: 100, 200 or up to 500 dollars, to try building tools.

How it workedtechnical detail

These figures are Anthropic's, from the launch post. On Terminal-Bench 4.0, a test of an agent working in a terminal, Haiku 5.5 scores 39.2 percent. Haiku 4.5 scored 0. GPT-6 Luna scores 16.4 percent. Sonnet 5.5 scores 70.6 percent. On OSWorld, using a computer to finish long tasks, Haiku scores 72.4 percent on the offline subset, against 15.7 for Haiku 4.5, 48.9 for Luna and 83.9 for Sonnet.

On knowledge work, Artificial Analysis's GDPval-AA, Haiku scores 1620 against 1437 for Luna and 1840 for Sonnet. It is the first Haiku with an effort dial, from low to max. More effort costs more.

Cyber safeguards are stricter than on Haiku 4.5 and looser than on Sonnet 5.5. Anthropic says they allow more defensive work than Sonnet does, and still block penetration testing. Biology safeguards match Sonnet 5, Sonnet 5.5 and Opus 5: research questions yes, requests judged likely to cause harm no. Wider access goes through the life-sciences program and the Cyber Verification Program.

A footnote: Haiku 5.5 is the fastest Claude at each model's ordinary speed. Opus in a separate fast mode is quicker.

What may follow

The model that runs a thousand small tasks a day is the one a company actually pays for. When that model gets cheap enough to drive a browser and a terminal, the agent incidents stop being a story about the largest model in a lab.

Sonnet is still the one Anthropic recommends for hard coding. Haiku is the one it recommends for the steps around it: summarise, look something up, click through a page. Those steps are most of the traffic.

What I do not know

I have not read the system card, only the launch post. The benchmarks are Anthropic's runs, including the numbers it prints for GPT-6 Luna. Artificial Analysis has not, in anything I have read, published its own Haiku 5.5 score yet. The Sonnet 5.5 entry still rests on Artificial Analysis, and those two sets of numbers are not the same test.

Customer quotes in the post, from Asana, HubSpot, Box and others, are the customers' and I have not checked them.

My notes

A 75 percent price cut on the small model is the story I would tell a Flemish company before I tell them about Astra. This is the one that will sit inside their support desk and their browser agent. The cyber block is looser than on the big model, and the computer-use score is no longer a toy.

Read next

Sources

  1. Anthropic: Introducing Claude Haiku 5.5primary · main source

    7 October 2026. Source for the price, the benchmarks, the safeguard description and the Sonnet cache cut. All comparison figures for other models are Anthropic's.